{
  "schema_version": 2,
  "generated_at": "2026-06-05T00:00:00Z",
  "bundle": "phase3-consortium-leworldmodel-evidence",
  "artifact_checks": [
    {
      "kind": "consortium-manifest",
      "label": "Phase 3 long-run consortium manifest",
      "location": "local",
      "uri": "docs/evidence/phase3_long_run_smoke_manifest.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "e20ba7853bb154353bf5f373028a9c06f3387610cbef425d0faa63ca44f20b6c",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "dataset-probe-registry",
      "label": "Phase 3 long-run dataset/probe registry",
      "location": "local",
      "uri": "docs/evidence/phase3_long_run_smoke_dataset_registry.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "2386c7a6cfdabcf584a12af1575f3bb42dbf4c5b23a1e0e72f0b2392ee66eacc",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "training-report",
      "label": "Phase 3 long-run training report",
      "location": "local",
      "uri": "docs/evidence/phase3_long_run_smoke_report.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "573c4248cabafe994478ae716f39fadd6d10ba009f64fa489add940f9abe77a0",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "privacy-aggregation-report",
      "label": "Phase 3 aggregation/privacy rows embedded in long-run report",
      "location": "local",
      "uri": "docs/evidence/phase3_long_run_smoke_report.json#rounds",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "573c4248cabafe994478ae716f39fadd6d10ba009f64fa489add940f9abe77a0",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "observability-report",
      "label": "Phase 3 observability/dropout report",
      "location": "local",
      "uri": "docs/evidence/phase3_observability_report.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "f944e50d7622fef1b4c0f279ae8c6d2957e5e1dcd1956329ece9b92635579375",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "eval-control-report",
      "label": "Phase 3 eval/control report",
      "location": "local",
      "uri": "docs/evidence/phase3_eval_report.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "ab6fcd1ed00ae0ed4f09e3bf5cf61f763cd913d23087f16ddf39923c2908dfdc",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "run-manifest",
      "label": "Phase 3 deterministic run manifest",
      "location": "local",
      "uri": "runs/phase3-long-run-smoke/phase3_run_manifest.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "cfaf14dcbe7c3fe9e64f2257729c047df448bd35218b387412f6a5d479c3169a",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "checkpoint-header",
      "label": "Phase 3 final checkpoint header",
      "location": "local",
      "uri": "runs/phase3-long-run-smoke/coordinator-artifacts/round-00010/header.json",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "00380f14da537080efa0969e212c93045be8768c443c59255420401635cec5b6",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    },
    {
      "kind": "checkpoint-weights",
      "label": "Phase 3 final checkpoint weights",
      "location": "local",
      "uri": "runs/phase3-long-run-smoke/coordinator-artifacts/round-00010/weights.safetensors",
      "checked_at": "2026-06-05T00:00:00Z",
      "exists": true,
      "sha256": "b669839b7c37b5f1185bc1326507a640d5d5991092135be1bbf2b4690b7754fd",
      "repo_type": null,
      "repo_id": null,
      "revision": null,
      "path_in_repo": null,
      "status_code": null,
      "error": null
    }
  ],
  "manifest": {
    "consortium_id": "lensemble-phase3-long-run-smoke",
    "run_id": "phase3-long-run-smoke-v1",
    "participant_count": 4,
    "coordinator_id": "phase3-long-run-coordinator",
    "protocol_version": "phase3-consortium-v1",
    "secure_aggregation_backend": "simulated",
    "dp_required": true,
    "public_probe_hash": "ddc8349fccfec07a41847e880f19574969d2a00de85a9353dd8c87cdeb7dfea2",
    "public_probe_hash_contract": "legacy-unscoped",
    "claim_boundary": "Historical pre-correctness-fix tiny in-process Phase 3 orchestration smoke. The declared secure_aggregation_required and dp_required fields describe the intended policy, not an achieved confidentiality guarantee: the optimizer consumed plaintext participant deltas, the simulated sum was a post-commit cross-check, and deterministic per-round noise snapshots were neither effective nor cumulative DP. The probe digest is legacy-unscoped. Audit history only; this artifact does not validate the corrected runtime or paper-scale robotics performance. Issue #335 tracks the rerun."
  },
  "dataset_registry": {
    "registry_id": "lensemble-phase3-long-run-smoke:phase3-long-run-smoke-v1:dataset-probe-registry",
    "run_mode": "public_example",
    "participant_count": 4,
    "participant_ids": [
      "phase3-so100-a",
      "phase3-so100-b",
      "phase3-so100-c",
      "phase3-so100-d"
    ],
    "public_probe_hash": "ddc8349fccfec07a41847e880f19574969d2a00de85a9353dd8c87cdeb7dfea2",
    "public_probe_hash_contract": "legacy-unscoped",
    "published_participant_count": 0,
    "placeholder_participant_count": 4,
    "raw_data_crosses_boundary": false,
    "heldout_policy": "last synthetic local episode held out"
  },
  "training": {
    "run_id": "phase3-long-run-smoke-v1",
    "evidence_status": "historical_pre_correctness_fix",
    "superseded_reason": "This run predates the correction of the outer-update direction and the public-probe-v2 target-binding contract. Its training and gauge metrics are retained for audit history only and do not validate the corrected runtime; rerun tracking is recorded in GitHub issue #335.",
    "participant_count": 4,
    "target_rounds": 10,
    "closed_rounds": 10,
    "completed_target": true,
    "config_hash": "a4f6bbabb919735c7679320c3f204cd0b1298e046e30cce8d11cbbbc8df07e57",
    "final_global_model_hash": "ed3081ee514af142a226443f113a37c24d7d5872bfb707f11abe10893a0ad50d",
    "run_manifest_hash": "cfaf14dcbe7c3fe9e64f2257729c047df448bd35218b387412f6a5d479c3169a",
    "checkpoint_header_sha256": "00380f14da537080efa0969e212c93045be8768c443c59255420401635cec5b6",
    "checkpoint_weights_sha256": "b669839b7c37b5f1185bc1326507a640d5d5991092135be1bbf2b4690b7754fd",
    "model_latent_dim": 8,
    "model_num_tokens": 4,
    "claim_boundary": "This run predates the correction of the outer-update direction and the public-probe-v2 target-binding contract. Its training and gauge metrics are retained for audit history only and do not validate the corrected runtime; rerun tracking is recorded in GitHub issue #335. Historical operational scope: the artifact records a deterministic tiny-model coordinator/participant lifecycle smoke. Its simulated sums were post-commit cross-checks over plaintext optimizer inputs, and its deterministic-noise epsilon values were one-round mechanism snapshots, not effective or cumulative DP."
  },
  "privacy_aggregation": {
    "secure_aggregation_backend": "simulated",
    "secure_aggregation_threshold": 3,
    "secure_sum_rounds": 0,
    "post_commit_cross_check_rounds": 10,
    "legacy_reported_secure_sum_rounds": 0,
    "dp_enabled": true,
    "dp_accountant": "rdp",
    "dp_epsilon": 8.0,
    "dp_delta": 1e-05,
    "dp_accounted_rounds": 10,
    "effective_dp_rounds": 0,
    "max_round_epsilon_spent": 5.302585092994046
  },
  "eval_controls": {
    "completed_controls": [
      "anchored-federation",
      "fork-a-frozen-encoder",
      "local-only",
      "naive-fedavg"
    ],
    "blocked_controls": [],
    "metric_row_count": 12,
    "blocked_task_env_ids": [
      "so100-heldout://phase3-public-task-scale"
    ],
    "model_card_eval_text": "Historical evidence status: these training and gauge rows predate the outer-update direction correction and the public-probe-v2 target-binding contract. They are retained for audit history only and do not validate the corrected runtime; rerun tracking is recorded in GitHub issue #335. Phase 3 evaluation evidence covers the local deterministic consortium-runtime smoke (participant-agent updates, 10 closed rounds, post-commit sum cross-check reporting, and one-round deterministic mechanism-accounting snapshots). Matched gauge controls are included only when their immutable reports and run manifests are supplied to the producer. Hash-bound historical gauge finding: anchored round-0 frame drift 48.9695 deg versus naive-FedAvg 180 deg; Fork-A 0 deg; local-only drift 180 deg with mean effective rank 120.316. These representation metrics are not robotics task-performance results. Historical limitation: at the old default outer step (outer_lr=0.7), the federated global representation collapsed over rounds (effective_rank -> 1). That trajectory also predates the outer-update correction, so it is failure-analysis history rather than a clean anchored-vs-naive result. Closed-loop task-scale SO-100 evaluation remains unrun; the published recorded splits cannot apply arbitrary planner actions to a live environment. Completed matched controls bound to published run hashes: anchored-federation, naive-fedavg, fork-a-frozen-encoder, local-only. These are representation-gauge controls and must not be described as completed robotics performance comparisons.",
    "claim_boundary": "Historical evidence status: these training and gauge rows predate the outer-update direction correction and the public-probe-v2 target-binding contract. They are retained for audit history only and do not validate the corrected runtime; rerun tracking is recorded in GitHub issue #335. Historical claim boundary: This Phase 3 eval report supports consortium-runtime engineering evidence from a local tiny-model run and explicit blockers for public task-scale controls. Post-commit sum cross-checks are not counted as secure aggregation, and deterministic-noise epsilon rows are one-round mechanism snapshots rather than effective DP. It does not claim paper-scale LeWorldModel performance or SO-100 robotics task success."
  },
  "observability": {
    "round_summary_count": 11,
    "dropout_decision_count": 1,
    "induced_dropout_outcomes": [
      "induced-dropout-close-with-quorum:closed"
    ],
    "redaction_contract_id": "phase3-observability-redaction-v1",
    "final_bundle_handoff": "Issue #230 must consume docs/evidence/phase3_observability_report.json in the final Phase 3 evidence bundle; this report includes an induced dropout event, so no no-failure exception is needed.",
    "claim_boundary": "Historical training/eval evidence: the referenced Phase 3 training and gauge rows predate the outer-update direction correction and the public-probe-v2 target-binding contract. They are retained for audit history only and do not validate the corrected runtime; rerun tracking is recorded in GitHub issue #335. Historical claim boundary: This Phase 3 observability report supports consortium-runtime, dropout, communication, artifact, and redaction evidence. It does not claim an optimizer-consumed secure sum or effective DP: the optimizer consumed plaintext participant updates, simulated sums are post-commit cross-checks, and epsilon values are one-round deterministic-replay snapshots. It does not claim paper-scale LeWorldModel performance or SO-100 robotics task success."
  },
  "publication": {
    "status": "historical",
    "model_repo_id": "abdelstark/lensemble-phase3-consortium-checkpoint",
    "model_repo_revision": "828e210cba4870b2be4ab573a5f0dd4ee30bae29",
    "dataset_repo_id": "abdelstark/lensemble-phase3-so100-silos",
    "dataset_repo_revision": "15f71911432b300dfdf41c998e27492e8c986be4",
    "dataset_bound_to_training": false,
    "expected_model_repo_files": [
      "README.md",
      "reports/phase3_evidence_bundle.json",
      "reports/phase3_long_run_smoke_report.json",
      "reports/phase3_eval_report.json",
      "reports/phase3_observability_report.json",
      "reports/phase3_long_run_manifest.json",
      "reports/phase3_long_run_dataset_registry.json",
      "artifacts/phase3_run_manifest.json",
      "artifacts/final/header.json",
      "artifacts/final/weights.safetensors"
    ],
    "blockers": [
      "The immutable Hub revisions are retained as historical audit artifacts; they do not contain evidence from the corrected outer update and target-bound probe contract.",
      "The associated SO-100 dataset revision is not bound to this historical synthetic-smoke training run.",
      "GitHub issue #335 tracks the corrected rerun and replacement publication."
    ]
  },
  "claim_boundaries": [
    "Historical evidence status: these Phase 3 training and gauge rows predate the outer-update direction correction and the public-probe-v2 target-binding contract. They are audit history and do not validate the corrected runtime; issue #335 tracks the rerun.",
    "Consortium-runtime evidence: four simulated participant agents completed a deterministic tiny-model lifecycle smoke with ten closed rounds; the optimizer consumed plaintext participant updates, while simulated sums were post-commit equivalence cross-checks.",
    "Training/eval scale: the bundled training run uses placeholder synthetic smoke data and tiny tokens/latent. Separately hash-bound control reports contain SO-100 representation metrics; neither surface is a paper-scale robotics result.",
    "Controls: anchored-federation, naive-FedAvg, Fork-A/frozen-encoder, and local-only controls are completed as hash-bound representation-metric rows; no matched control rows remain blocked.",
    "Privacy controls: deterministic noise and fresh one-round epsilon snapshots exercise mechanism plumbing but are not claim-grade effective DP or cumulative budget enforcement."
  ],
  "non_claims": [
    "The historical Phase 3 metrics are not evidence for the corrected outer update or the target-bound probe contract.",
    "Phase 3 does not include a provenance ledger implementation.",
    "Phase 3 does not cryptographically prove honest participant computation.",
    "Phase 3 does not claim paper-scale LeWorldModel performance.",
    "Phase 3 does not claim public SO-100 robotics task success.",
    "The bundle combines a synthetic runtime smoke with separately hash-bound historical real-data control reports; neither surface is a cryptographic honest-computation proof."
  ],
  "known_limitations": [
    "Corrected-runtime rerun (#335): every training/gauge metric in this bundle was produced before the outer-update sign and probe-target-binding fixes and must be regenerated before comparison or promotion.",
    "Secure aggregation is not integrated into the live optimizer path: current simulated/TEE sums are post-commit cross-checks over plaintext updates already consumed by the coordinator.",
    "Current DP noise is deterministically replayable from shared run configuration, and accounting is recreated per round; effective DP and cumulative enforcement are not claimed.",
    "DP-utility / federated-collapse (#244): the historical checkpoints exhibit global-representation collapse over rounds under the configured deterministic clipping/noise settings. This is a failure observation, not evidence of an effective or cumulative DP budget, and downstream planning success would be uninformative on these checkpoints.",
    "Downstream task-success (stable-worldmodel #96): closed-loop physical SO-100 task success is deferred, not claimed; it requires the unvendored stable-worldmodel planner suite and a non-collapsing federated checkpoint, because a recorded held-out split is open-loop and cannot apply arbitrary planner actions to recorded frames."
  ],
  "raw_data_in_report": false,
  "model_card_markdown": "---\nlicense: apache-2.0\nlibrary_name: lensemble\ntags:\n- federated-learning\n- world-model\n- jepa\n- robotics\n- phase3\n---\n\n# Lensemble Phase 3 Consortium JEPA World Model\n\nThis model repository preserves historical Phase 3 consortium-training audit evidence for a federated JEPA / LeWorldModel-flavour world model.\n\n\n## Historical Evidence Status\n\nThese training and gauge rows predate the outer-update direction correction and\nthe `public-probe-v2` target-binding contract. They are retained for audit\nhistory only and **do not validate the corrected runtime**. GitHub issue #335\ntracks the required rerun and replacement publication.\n\n\n## Consortium Runtime Evidence\n\n- Consortium id: `lensemble-phase3-long-run-smoke`\n- Run id: `phase3-long-run-smoke-v1`\n- Participant agents: 4\n- Coordinator: `phase3-long-run-coordinator`\n- Protocol: `phase3-consortium-v1`\n- Public probe hash: `ddc8349fccfec07a41847e880f19574969d2a00de85a9353dd8c87cdeb7dfea2`\n- Public probe hash contract: `legacy-unscoped`\n- Secure aggregation backend: `simulated`\n- DP accountant: `rdp`\n\n## Training And Evaluation Scale\n\n- Closed rounds: 10/10\n- Tiny model shape: `latent_dim=8`,\n  `num_tokens=4`\n- Config hash: `a4f6bbabb919735c7679320c3f204cd0b1298e046e30cce8d11cbbbc8df07e57`\n- Final checkpoint hash: `ed3081ee514af142a226443f113a37c24d7d5872bfb707f11abe10893a0ad50d`\n- Run-manifest hash: `cfaf14dcbe7c3fe9e64f2257729c047df448bd35218b387412f6a5d479c3169a`\n- Training evidence is a deterministic local consortium smoke, not a public\n  HF Jobs robotics-scale result.\n\n## Completed And Blocked Controls\n\n- Completed controls: `anchored-federation`, `fork-a-frozen-encoder`, `local-only`, `naive-fedavg`\n- Blocked controls: none\n- Eval/control metric rows: 12\n\nHistorical evidence status: these training and gauge rows predate the outer-update direction correction and the public-probe-v2 target-binding contract. They are retained for audit history only and do not validate the corrected runtime; rerun tracking is recorded in GitHub issue #335. Phase 3 evaluation evidence covers the local deterministic consortium-runtime smoke (participant-agent updates, 10 closed rounds, post-commit sum cross-check reporting, and one-round deterministic mechanism-accounting snapshots). Matched gauge controls are included only when their immutable reports and run manifests are supplied to the producer. Hash-bound historical gauge finding: anchored round-0 frame drift 48.9695 deg versus naive-FedAvg 180 deg; Fork-A 0 deg; local-only drift 180 deg with mean effective rank 120.316. These representation metrics are not robotics task-performance results. Historical limitation: at the old default outer step (outer_lr=0.7), the federated global representation collapsed over rounds (effective_rank -> 1). That trajectory also predates the outer-update correction, so it is failure-analysis history rather than a clean anchored-vs-naive result. Closed-loop task-scale SO-100 evaluation remains unrun; the published recorded splits cannot apply arbitrary planner actions to a live environment. Completed matched controls bound to published run hashes: anchored-federation, naive-fedavg, fork-a-frozen-encoder, local-only. These are representation-gauge controls and must not be described as completed robotics performance comparisons.\n\n## Privacy And Observability Controls\n\n- Optimizer-consumed secure-sum rounds: 0\n- Post-commit sum cross-check rounds: 10\n- One-round DP mechanism-accounting snapshots: 10\n- Claim-grade effective-DP rounds: 0\n- Max one-round epsilon snapshot: 5.302585092994046\n- Observability round summaries: 11\n- Induced dropout outcomes: induced-dropout-close-with-quorum:closed\n- Redaction contract: `phase3-observability-redaction-v1`\n\n## Dataset And Publication Status\n\n- Dataset registry: `lensemble-phase3-long-run-smoke:phase3-long-run-smoke-v1:dataset-probe-registry`\n- Dataset run mode: `public_example`\n- Participant data declarations: 4\n- Raw data crosses participant boundary: `False`\n- Model repo target: `hf://models/abdelstark/lensemble-phase3-consortium-checkpoint@828e210cba4870b2be4ab573a5f0dd4ee30bae29`\n- Dataset repo target: `hf://datasets/abdelstark/lensemble-phase3-so100-silos@15f71911432b300dfdf41c998e27492e8c986be4`\n- Dataset revision bound to this training run: `False`\n- Publication status: `historical`\n\n## Claim Boundaries\n\n- Historical evidence status: these Phase 3 training and gauge rows predate the outer-update direction correction and the public-probe-v2 target-binding contract. They are audit history and do not validate the corrected runtime; issue #335 tracks the rerun.\n- Consortium-runtime evidence: four simulated participant agents completed a deterministic tiny-model lifecycle smoke with ten closed rounds; the optimizer consumed plaintext participant updates, while simulated sums were post-commit equivalence cross-checks.\n- Training/eval scale: the bundled training run uses placeholder synthetic smoke data and tiny tokens/latent. Separately hash-bound control reports contain SO-100 representation metrics; neither surface is a paper-scale robotics result.\n- Controls: anchored-federation, naive-FedAvg, Fork-A/frozen-encoder, and local-only controls are completed as hash-bound representation-metric rows; no matched control rows remain blocked.\n- Privacy controls: deterministic noise and fresh one-round epsilon snapshots exercise mechanism plumbing but are not claim-grade effective DP or cumulative budget enforcement.\n\n## Non-Claims\n\n- The historical Phase 3 metrics are not evidence for the corrected outer update or the target-bound probe contract.\n- Phase 3 does not include a provenance ledger implementation.\n- Phase 3 does not cryptographically prove honest participant computation.\n- Phase 3 does not claim paper-scale LeWorldModel performance.\n- Phase 3 does not claim public SO-100 robotics task success.\n- The bundle combines a synthetic runtime smoke with separately hash-bound historical real-data control reports; neither surface is a cryptographic honest-computation proof.\n\n## Known Limitations\n\n- Corrected-runtime rerun (#335): every training/gauge metric in this bundle was produced before the outer-update sign and probe-target-binding fixes and must be regenerated before comparison or promotion.\n- Secure aggregation is not integrated into the live optimizer path: current simulated/TEE sums are post-commit cross-checks over plaintext updates already consumed by the coordinator.\n- Current DP noise is deterministically replayable from shared run configuration, and accounting is recreated per round; effective DP and cumulative enforcement are not claimed.\n- DP-utility / federated-collapse (#244): the historical checkpoints exhibit global-representation collapse over rounds under the configured deterministic clipping/noise settings. This is a failure observation, not evidence of an effective or cumulative DP budget, and downstream planning success would be uninformative on these checkpoints.\n- Downstream task-success (stable-worldmodel #96): closed-loop physical SO-100 task success is deferred, not claimed; it requires the unvendored stable-worldmodel planner suite and a non-collapsing federated checkpoint, because a recorded held-out split is open-loop and cannot apply arbitrary planner actions to recorded frames.\n\n## Reports In This Historical Record\n\n- `reports/phase3_evidence_bundle.json`\n- `reports/phase3_long_run_smoke_report.json`\n- `reports/phase3_eval_report.json`\n- `reports/phase3_observability_report.json`\n- `reports/phase3_long_run_smoke_manifest.json`\n- `reports/phase3_long_run_smoke_dataset_registry.json`\n- `artifacts/final/header.json`\n- `artifacts/final/weights.safetensors`\n"
}